URL
https://opencores.org/ocsvn/openrisc_me/openrisc_me/trunk
Subversion Repositories openrisc_me
[/] [openrisc/] [trunk/] [gnu-src/] [gcc-4.5.1/] [gcc/] [lambda-code.c] - Rev 295
Go to most recent revision | Compare with Previous | Blame | View Log
/* Loop transformation code generation Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. Contributed by Daniel Berlin <dberlin@dberlin.org> This file is part of GCC. GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version. GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see <http://www.gnu.org/licenses/>. */ #include "config.h" #include "system.h" #include "coretypes.h" #include "tm.h" #include "ggc.h" #include "tree.h" #include "target.h" #include "rtl.h" #include "basic-block.h" #include "diagnostic.h" #include "obstack.h" #include "tree-flow.h" #include "tree-dump.h" #include "timevar.h" #include "cfgloop.h" #include "expr.h" #include "optabs.h" #include "tree-chrec.h" #include "tree-data-ref.h" #include "tree-pass.h" #include "tree-scalar-evolution.h" #include "vec.h" #include "lambda.h" #include "vecprim.h" #include "pointer-set.h" /* This loop nest code generation is based on non-singular matrix math. A little terminology and a general sketch of the algorithm. See "A singular loop transformation framework based on non-singular matrices" by Wei Li and Keshav Pingali for formal proofs that the various statements below are correct. A loop iteration space represents the points traversed by the loop. A point in the iteration space can be represented by a vector of size <loop depth>. You can therefore represent the iteration space as an integral combinations of a set of basis vectors. A loop iteration space is dense if every integer point between the loop bounds is a point in the iteration space. Every loop with a step of 1 therefore has a dense iteration space. for i = 1 to 3, step 1 is a dense iteration space. A loop iteration space is sparse if it is not dense. That is, the iteration space skips integer points that are within the loop bounds. for i = 1 to 3, step 2 is a sparse iteration space, because the integer point 2 is skipped. Dense source spaces are easy to transform, because they don't skip any points to begin with. Thus we can compute the exact bounds of the target space using min/max and floor/ceil. For a dense source space, we take the transformation matrix, decompose it into a lower triangular part (H) and a unimodular part (U). We then compute the auxiliary space from the unimodular part (source loop nest . U = auxiliary space) , which has two important properties: 1. It traverses the iterations in the same lexicographic order as the source space. 2. It is a dense space when the source is a dense space (even if the target space is going to be sparse). Given the auxiliary space, we use the lower triangular part to compute the bounds in the target space by simple matrix multiplication. The gaps in the target space (IE the new loop step sizes) will be the diagonals of the H matrix. Sparse source spaces require another step, because you can't directly compute the exact bounds of the auxiliary and target space from the sparse space. Rather than try to come up with a separate algorithm to handle sparse source spaces directly, we just find a legal transformation matrix that gives you the sparse source space, from a dense space, and then transform the dense space. For a regular sparse space, you can represent the source space as an integer lattice, and the base space of that lattice will always be dense. Thus, we effectively use the lattice to figure out the transformation from the lattice base space, to the sparse iteration space (IE what transform was applied to the dense space to make it sparse). We then compose this transform with the transformation matrix specified by the user (since our matrix transformations are closed under composition, this is okay). We can then use the base space (which is dense) plus the composed transformation matrix, to compute the rest of the transform using the dense space algorithm above. In other words, our sparse source space (B) is decomposed into a dense base space (A), and a matrix (L) that transforms A into B, such that A.L = B. We then compute the composition of L and the user transformation matrix (T), so that T is now a transform from A to the result, instead of from B to the result. IE A.(LT) = result instead of B.T = result Since A is now a dense source space, we can use the dense source space algorithm above to compute the result of applying transform (LT) to A. Fourier-Motzkin elimination is used to compute the bounds of the base space of the lattice. */ static bool perfect_nestify (struct loop *, VEC(tree,heap) *, VEC(tree,heap) *, VEC(int,heap) *, VEC(tree,heap) *); /* Lattice stuff that is internal to the code generation algorithm. */ typedef struct lambda_lattice_s { /* Lattice base matrix. */ lambda_matrix base; /* Lattice dimension. */ int dimension; /* Origin vector for the coefficients. */ lambda_vector origin; /* Origin matrix for the invariants. */ lambda_matrix origin_invariants; /* Number of invariants. */ int invariants; } *lambda_lattice; #define LATTICE_BASE(T) ((T)->base) #define LATTICE_DIMENSION(T) ((T)->dimension) #define LATTICE_ORIGIN(T) ((T)->origin) #define LATTICE_ORIGIN_INVARIANTS(T) ((T)->origin_invariants) #define LATTICE_INVARIANTS(T) ((T)->invariants) static bool lle_equal (lambda_linear_expression, lambda_linear_expression, int, int); static lambda_lattice lambda_lattice_new (int, int, struct obstack *); static lambda_lattice lambda_lattice_compute_base (lambda_loopnest, struct obstack *); static bool can_convert_to_perfect_nest (struct loop *); /* Create a new lambda body vector. */ lambda_body_vector lambda_body_vector_new (int size, struct obstack * lambda_obstack) { lambda_body_vector ret; ret = (lambda_body_vector)obstack_alloc (lambda_obstack, sizeof (*ret)); LBV_COEFFICIENTS (ret) = lambda_vector_new (size); LBV_SIZE (ret) = size; LBV_DENOMINATOR (ret) = 1; return ret; } /* Compute the new coefficients for the vector based on the *inverse* of the transformation matrix. */ lambda_body_vector lambda_body_vector_compute_new (lambda_trans_matrix transform, lambda_body_vector vect, struct obstack * lambda_obstack) { lambda_body_vector temp; int depth; /* Make sure the matrix is square. */ gcc_assert (LTM_ROWSIZE (transform) == LTM_COLSIZE (transform)); depth = LTM_ROWSIZE (transform); temp = lambda_body_vector_new (depth, lambda_obstack); LBV_DENOMINATOR (temp) = LBV_DENOMINATOR (vect) * LTM_DENOMINATOR (transform); lambda_vector_matrix_mult (LBV_COEFFICIENTS (vect), depth, LTM_MATRIX (transform), depth, LBV_COEFFICIENTS (temp)); LBV_SIZE (temp) = LBV_SIZE (vect); return temp; } /* Print out a lambda body vector. */ void print_lambda_body_vector (FILE * outfile, lambda_body_vector body) { print_lambda_vector (outfile, LBV_COEFFICIENTS (body), LBV_SIZE (body)); } /* Return TRUE if two linear expressions are equal. */ static bool lle_equal (lambda_linear_expression lle1, lambda_linear_expression lle2, int depth, int invariants) { int i; if (lle1 == NULL || lle2 == NULL) return false; if (LLE_CONSTANT (lle1) != LLE_CONSTANT (lle2)) return false; if (LLE_DENOMINATOR (lle1) != LLE_DENOMINATOR (lle2)) return false; for (i = 0; i < depth; i++) if (LLE_COEFFICIENTS (lle1)[i] != LLE_COEFFICIENTS (lle2)[i]) return false; for (i = 0; i < invariants; i++) if (LLE_INVARIANT_COEFFICIENTS (lle1)[i] != LLE_INVARIANT_COEFFICIENTS (lle2)[i]) return false; return true; } /* Create a new linear expression with dimension DIM, and total number of invariants INVARIANTS. */ lambda_linear_expression lambda_linear_expression_new (int dim, int invariants, struct obstack * lambda_obstack) { lambda_linear_expression ret; ret = (lambda_linear_expression)obstack_alloc (lambda_obstack, sizeof (*ret)); LLE_COEFFICIENTS (ret) = lambda_vector_new (dim); LLE_CONSTANT (ret) = 0; LLE_INVARIANT_COEFFICIENTS (ret) = lambda_vector_new (invariants); LLE_DENOMINATOR (ret) = 1; LLE_NEXT (ret) = NULL; return ret; } /* Print out a linear expression EXPR, with SIZE coefficients, to OUTFILE. The starting letter used for variable names is START. */ static void print_linear_expression (FILE * outfile, lambda_vector expr, int size, char start) { int i; bool first = true; for (i = 0; i < size; i++) { if (expr[i] != 0) { if (first) { if (expr[i] < 0) fprintf (outfile, "-"); first = false; } else if (expr[i] > 0) fprintf (outfile, " + "); else fprintf (outfile, " - "); if (abs (expr[i]) == 1) fprintf (outfile, "%c", start + i); else fprintf (outfile, "%d%c", abs (expr[i]), start + i); } } } /* Print out a lambda linear expression structure, EXPR, to OUTFILE. The depth/number of coefficients is given by DEPTH, the number of invariants is given by INVARIANTS, and the character to start variable names with is given by START. */ void print_lambda_linear_expression (FILE * outfile, lambda_linear_expression expr, int depth, int invariants, char start) { fprintf (outfile, "\tLinear expression: "); print_linear_expression (outfile, LLE_COEFFICIENTS (expr), depth, start); fprintf (outfile, " constant: %d ", LLE_CONSTANT (expr)); fprintf (outfile, " invariants: "); print_linear_expression (outfile, LLE_INVARIANT_COEFFICIENTS (expr), invariants, 'A'); fprintf (outfile, " denominator: %d\n", LLE_DENOMINATOR (expr)); } /* Print a lambda loop structure LOOP to OUTFILE. The depth/number of coefficients is given by DEPTH, the number of invariants is given by INVARIANTS, and the character to start variable names with is given by START. */ void print_lambda_loop (FILE * outfile, lambda_loop loop, int depth, int invariants, char start) { int step; lambda_linear_expression expr; gcc_assert (loop); expr = LL_LINEAR_OFFSET (loop); step = LL_STEP (loop); fprintf (outfile, " step size = %d \n", step); if (expr) { fprintf (outfile, " linear offset: \n"); print_lambda_linear_expression (outfile, expr, depth, invariants, start); } fprintf (outfile, " lower bound: \n"); for (expr = LL_LOWER_BOUND (loop); expr != NULL; expr = LLE_NEXT (expr)) print_lambda_linear_expression (outfile, expr, depth, invariants, start); fprintf (outfile, " upper bound: \n"); for (expr = LL_UPPER_BOUND (loop); expr != NULL; expr = LLE_NEXT (expr)) print_lambda_linear_expression (outfile, expr, depth, invariants, start); } /* Create a new loop nest structure with DEPTH loops, and INVARIANTS as the number of invariants. */ lambda_loopnest lambda_loopnest_new (int depth, int invariants, struct obstack * lambda_obstack) { lambda_loopnest ret; ret = (lambda_loopnest)obstack_alloc (lambda_obstack, sizeof (*ret)); LN_LOOPS (ret) = (lambda_loop *) obstack_alloc (lambda_obstack, depth * sizeof(LN_LOOPS(ret))); LN_DEPTH (ret) = depth; LN_INVARIANTS (ret) = invariants; return ret; } /* Print a lambda loopnest structure, NEST, to OUTFILE. The starting character to use for loop names is given by START. */ void print_lambda_loopnest (FILE * outfile, lambda_loopnest nest, char start) { int i; for (i = 0; i < LN_DEPTH (nest); i++) { fprintf (outfile, "Loop %c\n", start + i); print_lambda_loop (outfile, LN_LOOPS (nest)[i], LN_DEPTH (nest), LN_INVARIANTS (nest), 'i'); fprintf (outfile, "\n"); } } /* Allocate a new lattice structure of DEPTH x DEPTH, with INVARIANTS number of invariants. */ static lambda_lattice lambda_lattice_new (int depth, int invariants, struct obstack * lambda_obstack) { lambda_lattice ret = (lambda_lattice)obstack_alloc (lambda_obstack, sizeof (*ret)); LATTICE_BASE (ret) = lambda_matrix_new (depth, depth); LATTICE_ORIGIN (ret) = lambda_vector_new (depth); LATTICE_ORIGIN_INVARIANTS (ret) = lambda_matrix_new (depth, invariants); LATTICE_DIMENSION (ret) = depth; LATTICE_INVARIANTS (ret) = invariants; return ret; } /* Compute the lattice base for NEST. The lattice base is essentially a non-singular transform from a dense base space to a sparse iteration space. We use it so that we don't have to specially handle the case of a sparse iteration space in other parts of the algorithm. As a result, this routine only does something interesting (IE produce a matrix that isn't the identity matrix) if NEST is a sparse space. */ static lambda_lattice lambda_lattice_compute_base (lambda_loopnest nest, struct obstack * lambda_obstack) { lambda_lattice ret; int depth, invariants; lambda_matrix base; int i, j, step; lambda_loop loop; lambda_linear_expression expression; depth = LN_DEPTH (nest); invariants = LN_INVARIANTS (nest); ret = lambda_lattice_new (depth, invariants, lambda_obstack); base = LATTICE_BASE (ret); for (i = 0; i < depth; i++) { loop = LN_LOOPS (nest)[i]; gcc_assert (loop); step = LL_STEP (loop); /* If we have a step of 1, then the base is one, and the origin and invariant coefficients are 0. */ if (step == 1) { for (j = 0; j < depth; j++) base[i][j] = 0; base[i][i] = 1; LATTICE_ORIGIN (ret)[i] = 0; for (j = 0; j < invariants; j++) LATTICE_ORIGIN_INVARIANTS (ret)[i][j] = 0; } else { /* Otherwise, we need the lower bound expression (which must be an affine function) to determine the base. */ expression = LL_LOWER_BOUND (loop); gcc_assert (expression && !LLE_NEXT (expression) && LLE_DENOMINATOR (expression) == 1); /* The lower triangular portion of the base is going to be the coefficient times the step */ for (j = 0; j < i; j++) base[i][j] = LLE_COEFFICIENTS (expression)[j] * LL_STEP (LN_LOOPS (nest)[j]); base[i][i] = step; for (j = i + 1; j < depth; j++) base[i][j] = 0; /* Origin for this loop is the constant of the lower bound expression. */ LATTICE_ORIGIN (ret)[i] = LLE_CONSTANT (expression); /* Coefficient for the invariants are equal to the invariant coefficients in the expression. */ for (j = 0; j < invariants; j++) LATTICE_ORIGIN_INVARIANTS (ret)[i][j] = LLE_INVARIANT_COEFFICIENTS (expression)[j]; } } return ret; } /* Compute the least common multiple of two numbers A and B . */ int least_common_multiple (int a, int b) { return (abs (a) * abs (b) / gcd (a, b)); } /* Perform Fourier-Motzkin elimination to calculate the bounds of the auxiliary nest. Fourier-Motzkin is a way of reducing systems of linear inequalities so that it is easy to calculate the answer and bounds. A sketch of how it works: Given a system of linear inequalities, ai * xj >= bk, you can always rewrite the constraints so they are all of the form a <= x, or x <= b, or x >= constant for some x in x1 ... xj (and some b in b1 ... bk, and some a in a1...ai) You can then eliminate this x from the non-constant inequalities by rewriting these as a <= b, x >= constant, and delete the x variable. You can then repeat this for any remaining x variables, and then we have an easy to use variable <= constant (or no variables at all) form that we can construct our bounds from. In our case, each time we eliminate, we construct part of the bound from the ith variable, then delete the ith variable. Remember the constant are in our vector a, our coefficient matrix is A, and our invariant coefficient matrix is B. SIZE is the size of the matrices being passed. DEPTH is the loop nest depth. INVARIANTS is the number of loop invariants. A, B, and a are the coefficient matrix, invariant coefficient, and a vector of constants, respectively. */ static lambda_loopnest compute_nest_using_fourier_motzkin (int size, int depth, int invariants, lambda_matrix A, lambda_matrix B, lambda_vector a, struct obstack * lambda_obstack) { int multiple, f1, f2; int i, j, k; lambda_linear_expression expression; lambda_loop loop; lambda_loopnest auxillary_nest; lambda_matrix swapmatrix, A1, B1; lambda_vector swapvector, a1; int newsize; A1 = lambda_matrix_new (128, depth); B1 = lambda_matrix_new (128, invariants); a1 = lambda_vector_new (128); auxillary_nest = lambda_loopnest_new (depth, invariants, lambda_obstack); for (i = depth - 1; i >= 0; i--) { loop = lambda_loop_new (); LN_LOOPS (auxillary_nest)[i] = loop; LL_STEP (loop) = 1; for (j = 0; j < size; j++) { if (A[j][i] < 0) { /* Any linear expression in the matrix with a coefficient less than 0 becomes part of the new lower bound. */ expression = lambda_linear_expression_new (depth, invariants, lambda_obstack); for (k = 0; k < i; k++) LLE_COEFFICIENTS (expression)[k] = A[j][k]; for (k = 0; k < invariants; k++) LLE_INVARIANT_COEFFICIENTS (expression)[k] = -1 * B[j][k]; LLE_DENOMINATOR (expression) = -1 * A[j][i]; LLE_CONSTANT (expression) = -1 * a[j]; /* Ignore if identical to the existing lower bound. */ if (!lle_equal (LL_LOWER_BOUND (loop), expression, depth, invariants)) { LLE_NEXT (expression) = LL_LOWER_BOUND (loop); LL_LOWER_BOUND (loop) = expression; } } else if (A[j][i] > 0) { /* Any linear expression with a coefficient greater than 0 becomes part of the new upper bound. */ expression = lambda_linear_expression_new (depth, invariants, lambda_obstack); for (k = 0; k < i; k++) LLE_COEFFICIENTS (expression)[k] = -1 * A[j][k]; for (k = 0; k < invariants; k++) LLE_INVARIANT_COEFFICIENTS (expression)[k] = B[j][k]; LLE_DENOMINATOR (expression) = A[j][i]; LLE_CONSTANT (expression) = a[j]; /* Ignore if identical to the existing upper bound. */ if (!lle_equal (LL_UPPER_BOUND (loop), expression, depth, invariants)) { LLE_NEXT (expression) = LL_UPPER_BOUND (loop); LL_UPPER_BOUND (loop) = expression; } } } /* This portion creates a new system of linear inequalities by deleting the i'th variable, reducing the system by one variable. */ newsize = 0; for (j = 0; j < size; j++) { /* If the coefficient for the i'th variable is 0, then we can just eliminate the variable straightaway. Otherwise, we have to multiply through by the coefficients we are eliminating. */ if (A[j][i] == 0) { lambda_vector_copy (A[j], A1[newsize], depth); lambda_vector_copy (B[j], B1[newsize], invariants); a1[newsize] = a[j]; newsize++; } else if (A[j][i] > 0) { for (k = 0; k < size; k++) { if (A[k][i] < 0) { multiple = least_common_multiple (A[j][i], A[k][i]); f1 = multiple / A[j][i]; f2 = -1 * multiple / A[k][i]; lambda_vector_add_mc (A[j], f1, A[k], f2, A1[newsize], depth); lambda_vector_add_mc (B[j], f1, B[k], f2, B1[newsize], invariants); a1[newsize] = f1 * a[j] + f2 * a[k]; newsize++; } } } } swapmatrix = A; A = A1; A1 = swapmatrix; swapmatrix = B; B = B1; B1 = swapmatrix; swapvector = a; a = a1; a1 = swapvector; size = newsize; } return auxillary_nest; } /* Compute the loop bounds for the auxiliary space NEST. Input system used is Ax <= b. TRANS is the unimodular transformation. Given the original nest, this function will 1. Convert the nest into matrix form, which consists of a matrix for the coefficients, a matrix for the invariant coefficients, and a vector for the constants. 2. Use the matrix form to calculate the lattice base for the nest (which is a dense space) 3. Compose the dense space transform with the user specified transform, to get a transform we can easily calculate transformed bounds for. 4. Multiply the composed transformation matrix times the matrix form of the loop. 5. Transform the newly created matrix (from step 4) back into a loop nest using Fourier-Motzkin elimination to figure out the bounds. */ static lambda_loopnest lambda_compute_auxillary_space (lambda_loopnest nest, lambda_trans_matrix trans, struct obstack * lambda_obstack) { lambda_matrix A, B, A1, B1; lambda_vector a, a1; lambda_matrix invertedtrans; int depth, invariants, size; int i, j; lambda_loop loop; lambda_linear_expression expression; lambda_lattice lattice; depth = LN_DEPTH (nest); invariants = LN_INVARIANTS (nest); /* Unfortunately, we can't know the number of constraints we'll have ahead of time, but this should be enough even in ridiculous loop nest cases. We must not go over this limit. */ A = lambda_matrix_new (128, depth); B = lambda_matrix_new (128, invariants); a = lambda_vector_new (128); A1 = lambda_matrix_new (128, depth); B1 = lambda_matrix_new (128, invariants); a1 = lambda_vector_new (128); /* Store the bounds in the equation matrix A, constant vector a, and invariant matrix B, so that we have Ax <= a + B. This requires a little equation rearranging so that everything is on the correct side of the inequality. */ size = 0; for (i = 0; i < depth; i++) { loop = LN_LOOPS (nest)[i]; /* First we do the lower bound. */ if (LL_STEP (loop) > 0) expression = LL_LOWER_BOUND (loop); else expression = LL_UPPER_BOUND (loop); for (; expression != NULL; expression = LLE_NEXT (expression)) { /* Fill in the coefficient. */ for (j = 0; j < i; j++) A[size][j] = LLE_COEFFICIENTS (expression)[j]; /* And the invariant coefficient. */ for (j = 0; j < invariants; j++) B[size][j] = LLE_INVARIANT_COEFFICIENTS (expression)[j]; /* And the constant. */ a[size] = LLE_CONSTANT (expression); /* Convert (2x+3y+2+b)/4 <= z to 2x+3y-4z <= -2-b. IE put all constants and single variables on */ A[size][i] = -1 * LLE_DENOMINATOR (expression); a[size] *= -1; for (j = 0; j < invariants; j++) B[size][j] *= -1; size++; /* Need to increase matrix sizes above. */ gcc_assert (size <= 127); } /* Then do the exact same thing for the upper bounds. */ if (LL_STEP (loop) > 0) expression = LL_UPPER_BOUND (loop); else expression = LL_LOWER_BOUND (loop); for (; expression != NULL; expression = LLE_NEXT (expression)) { /* Fill in the coefficient. */ for (j = 0; j < i; j++) A[size][j] = LLE_COEFFICIENTS (expression)[j]; /* And the invariant coefficient. */ for (j = 0; j < invariants; j++) B[size][j] = LLE_INVARIANT_COEFFICIENTS (expression)[j]; /* And the constant. */ a[size] = LLE_CONSTANT (expression); /* Convert z <= (2x+3y+2+b)/4 to -2x-3y+4z <= 2+b. */ for (j = 0; j < i; j++) A[size][j] *= -1; A[size][i] = LLE_DENOMINATOR (expression); size++; /* Need to increase matrix sizes above. */ gcc_assert (size <= 127); } } /* Compute the lattice base x = base * y + origin, where y is the base space. */ lattice = lambda_lattice_compute_base (nest, lambda_obstack); /* Ax <= a + B then becomes ALy <= a+B - A*origin. L is the lattice base */ /* A1 = A * L */ lambda_matrix_mult (A, LATTICE_BASE (lattice), A1, size, depth, depth); /* a1 = a - A * origin constant. */ lambda_matrix_vector_mult (A, size, depth, LATTICE_ORIGIN (lattice), a1); lambda_vector_add_mc (a, 1, a1, -1, a1, size); /* B1 = B - A * origin invariant. */ lambda_matrix_mult (A, LATTICE_ORIGIN_INVARIANTS (lattice), B1, size, depth, invariants); lambda_matrix_add_mc (B, 1, B1, -1, B1, size, invariants); /* Now compute the auxiliary space bounds by first inverting U, multiplying it by A1, then performing Fourier-Motzkin. */ invertedtrans = lambda_matrix_new (depth, depth); /* Compute the inverse of U. */ lambda_matrix_inverse (LTM_MATRIX (trans), invertedtrans, depth); /* A = A1 inv(U). */ lambda_matrix_mult (A1, invertedtrans, A, size, depth, depth); return compute_nest_using_fourier_motzkin (size, depth, invariants, A, B1, a1, lambda_obstack); } /* Compute the loop bounds for the target space, using the bounds of the auxiliary nest AUXILLARY_NEST, and the triangular matrix H. The target space loop bounds are computed by multiplying the triangular matrix H by the auxiliary nest, to get the new loop bounds. The sign of the loop steps (positive or negative) is then used to swap the bounds if the loop counts downwards. Return the target loopnest. */ static lambda_loopnest lambda_compute_target_space (lambda_loopnest auxillary_nest, lambda_trans_matrix H, lambda_vector stepsigns, struct obstack * lambda_obstack) { lambda_matrix inverse, H1; int determinant, i, j; int gcd1, gcd2; int factor; lambda_loopnest target_nest; int depth, invariants; lambda_matrix target; lambda_loop auxillary_loop, target_loop; lambda_linear_expression expression, auxillary_expr, target_expr, tmp_expr; depth = LN_DEPTH (auxillary_nest); invariants = LN_INVARIANTS (auxillary_nest); inverse = lambda_matrix_new (depth, depth); determinant = lambda_matrix_inverse (LTM_MATRIX (H), inverse, depth); /* H1 is H excluding its diagonal. */ H1 = lambda_matrix_new (depth, depth); lambda_matrix_copy (LTM_MATRIX (H), H1, depth, depth); for (i = 0; i < depth; i++) H1[i][i] = 0; /* Computes the linear offsets of the loop bounds. */ target = lambda_matrix_new (depth, depth); lambda_matrix_mult (H1, inverse, target, depth, depth, depth); target_nest = lambda_loopnest_new (depth, invariants, lambda_obstack); for (i = 0; i < depth; i++) { /* Get a new loop structure. */ target_loop = lambda_loop_new (); LN_LOOPS (target_nest)[i] = target_loop; /* Computes the gcd of the coefficients of the linear part. */ gcd1 = lambda_vector_gcd (target[i], i); /* Include the denominator in the GCD. */ gcd1 = gcd (gcd1, determinant); /* Now divide through by the gcd. */ for (j = 0; j < i; j++) target[i][j] = target[i][j] / gcd1; expression = lambda_linear_expression_new (depth, invariants, lambda_obstack); lambda_vector_copy (target[i], LLE_COEFFICIENTS (expression), depth); LLE_DENOMINATOR (expression) = determinant / gcd1; LLE_CONSTANT (expression) = 0; lambda_vector_clear (LLE_INVARIANT_COEFFICIENTS (expression), invariants); LL_LINEAR_OFFSET (target_loop) = expression; } /* For each loop, compute the new bounds from H. */ for (i = 0; i < depth; i++) { auxillary_loop = LN_LOOPS (auxillary_nest)[i]; target_loop = LN_LOOPS (target_nest)[i]; LL_STEP (target_loop) = LTM_MATRIX (H)[i][i]; factor = LTM_MATRIX (H)[i][i]; /* First we do the lower bound. */ auxillary_expr = LL_LOWER_BOUND (auxillary_loop); for (; auxillary_expr != NULL; auxillary_expr = LLE_NEXT (auxillary_expr)) { target_expr = lambda_linear_expression_new (depth, invariants, lambda_obstack); lambda_vector_matrix_mult (LLE_COEFFICIENTS (auxillary_expr), depth, inverse, depth, LLE_COEFFICIENTS (target_expr)); lambda_vector_mult_const (LLE_COEFFICIENTS (target_expr), LLE_COEFFICIENTS (target_expr), depth, factor); LLE_CONSTANT (target_expr) = LLE_CONSTANT (auxillary_expr) * factor; lambda_vector_copy (LLE_INVARIANT_COEFFICIENTS (auxillary_expr), LLE_INVARIANT_COEFFICIENTS (target_expr), invariants); lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS (target_expr), LLE_INVARIANT_COEFFICIENTS (target_expr), invariants, factor); LLE_DENOMINATOR (target_expr) = LLE_DENOMINATOR (auxillary_expr); if (!lambda_vector_zerop (LLE_COEFFICIENTS (target_expr), depth)) { LLE_CONSTANT (target_expr) = LLE_CONSTANT (target_expr) * determinant; lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS (target_expr), LLE_INVARIANT_COEFFICIENTS (target_expr), invariants, determinant); LLE_DENOMINATOR (target_expr) = LLE_DENOMINATOR (target_expr) * determinant; } /* Find the gcd and divide by it here, rather than doing it at the tree level. */ gcd1 = lambda_vector_gcd (LLE_COEFFICIENTS (target_expr), depth); gcd2 = lambda_vector_gcd (LLE_INVARIANT_COEFFICIENTS (target_expr), invariants); gcd1 = gcd (gcd1, gcd2); gcd1 = gcd (gcd1, LLE_CONSTANT (target_expr)); gcd1 = gcd (gcd1, LLE_DENOMINATOR (target_expr)); for (j = 0; j < depth; j++) LLE_COEFFICIENTS (target_expr)[j] /= gcd1; for (j = 0; j < invariants; j++) LLE_INVARIANT_COEFFICIENTS (target_expr)[j] /= gcd1; LLE_CONSTANT (target_expr) /= gcd1; LLE_DENOMINATOR (target_expr) /= gcd1; /* Ignore if identical to existing bound. */ if (!lle_equal (LL_LOWER_BOUND (target_loop), target_expr, depth, invariants)) { LLE_NEXT (target_expr) = LL_LOWER_BOUND (target_loop); LL_LOWER_BOUND (target_loop) = target_expr; } } /* Now do the upper bound. */ auxillary_expr = LL_UPPER_BOUND (auxillary_loop); for (; auxillary_expr != NULL; auxillary_expr = LLE_NEXT (auxillary_expr)) { target_expr = lambda_linear_expression_new (depth, invariants, lambda_obstack); lambda_vector_matrix_mult (LLE_COEFFICIENTS (auxillary_expr), depth, inverse, depth, LLE_COEFFICIENTS (target_expr)); lambda_vector_mult_const (LLE_COEFFICIENTS (target_expr), LLE_COEFFICIENTS (target_expr), depth, factor); LLE_CONSTANT (target_expr) = LLE_CONSTANT (auxillary_expr) * factor; lambda_vector_copy (LLE_INVARIANT_COEFFICIENTS (auxillary_expr), LLE_INVARIANT_COEFFICIENTS (target_expr), invariants); lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS (target_expr), LLE_INVARIANT_COEFFICIENTS (target_expr), invariants, factor); LLE_DENOMINATOR (target_expr) = LLE_DENOMINATOR (auxillary_expr); if (!lambda_vector_zerop (LLE_COEFFICIENTS (target_expr), depth)) { LLE_CONSTANT (target_expr) = LLE_CONSTANT (target_expr) * determinant; lambda_vector_mult_const (LLE_INVARIANT_COEFFICIENTS (target_expr), LLE_INVARIANT_COEFFICIENTS (target_expr), invariants, determinant); LLE_DENOMINATOR (target_expr) = LLE_DENOMINATOR (target_expr) * determinant; } /* Find the gcd and divide by it here, instead of at the tree level. */ gcd1 = lambda_vector_gcd (LLE_COEFFICIENTS (target_expr), depth); gcd2 = lambda_vector_gcd (LLE_INVARIANT_COEFFICIENTS (target_expr), invariants); gcd1 = gcd (gcd1, gcd2); gcd1 = gcd (gcd1, LLE_CONSTANT (target_expr)); gcd1 = gcd (gcd1, LLE_DENOMINATOR (target_expr)); for (j = 0; j < depth; j++) LLE_COEFFICIENTS (target_expr)[j] /= gcd1; for (j = 0; j < invariants; j++) LLE_INVARIANT_COEFFICIENTS (target_expr)[j] /= gcd1; LLE_CONSTANT (target_expr) /= gcd1; LLE_DENOMINATOR (target_expr) /= gcd1; /* Ignore if equal to existing bound. */ if (!lle_equal (LL_UPPER_BOUND (target_loop), target_expr, depth, invariants)) { LLE_NEXT (target_expr) = LL_UPPER_BOUND (target_loop); LL_UPPER_BOUND (target_loop) = target_expr; } } } for (i = 0; i < depth; i++) { target_loop = LN_LOOPS (target_nest)[i]; /* If necessary, exchange the upper and lower bounds and negate the step size. */ if (stepsigns[i] < 0) { LL_STEP (target_loop) *= -1; tmp_expr = LL_LOWER_BOUND (target_loop); LL_LOWER_BOUND (target_loop) = LL_UPPER_BOUND (target_loop); LL_UPPER_BOUND (target_loop) = tmp_expr; } } return target_nest; } /* Compute the step signs of TRANS, using TRANS and stepsigns. Return the new result. */ static lambda_vector lambda_compute_step_signs (lambda_trans_matrix trans, lambda_vector stepsigns) { lambda_matrix matrix, H; int size; lambda_vector newsteps; int i, j, factor, minimum_column; int temp; matrix = LTM_MATRIX (trans); size = LTM_ROWSIZE (trans); H = lambda_matrix_new (size, size); newsteps = lambda_vector_new (size); lambda_vector_copy (stepsigns, newsteps, size); lambda_matrix_copy (matrix, H, size, size); for (j = 0; j < size; j++) { lambda_vector row; row = H[j]; for (i = j; i < size; i++) if (row[i] < 0) lambda_matrix_col_negate (H, size, i); while (lambda_vector_first_nz (row, size, j + 1) < size) { minimum_column = lambda_vector_min_nz (row, size, j); lambda_matrix_col_exchange (H, size, j, minimum_column); temp = newsteps[j]; newsteps[j] = newsteps[minimum_column]; newsteps[minimum_column] = temp; for (i = j + 1; i < size; i++) { factor = row[i] / row[j]; lambda_matrix_col_add (H, size, j, i, -1 * factor); } } } return newsteps; } /* Transform NEST according to TRANS, and return the new loopnest. This involves 1. Computing a lattice base for the transformation 2. Composing the dense base with the specified transformation (TRANS) 3. Decomposing the combined transformation into a lower triangular portion, and a unimodular portion. 4. Computing the auxiliary nest using the unimodular portion. 5. Computing the target nest using the auxiliary nest and the lower triangular portion. */ lambda_loopnest lambda_loopnest_transform (lambda_loopnest nest, lambda_trans_matrix trans, struct obstack * lambda_obstack) { lambda_loopnest auxillary_nest, target_nest; int depth, invariants; int i, j; lambda_lattice lattice; lambda_trans_matrix trans1, H, U; lambda_loop loop; lambda_linear_expression expression; lambda_vector origin; lambda_matrix origin_invariants; lambda_vector stepsigns; int f; depth = LN_DEPTH (nest); invariants = LN_INVARIANTS (nest); /* Keep track of the signs of the loop steps. */ stepsigns = lambda_vector_new (depth); for (i = 0; i < depth; i++) { if (LL_STEP (LN_LOOPS (nest)[i]) > 0) stepsigns[i] = 1; else stepsigns[i] = -1; } /* Compute the lattice base. */ lattice = lambda_lattice_compute_base (nest, lambda_obstack); trans1 = lambda_trans_matrix_new (depth, depth); /* Multiply the transformation matrix by the lattice base. */ lambda_matrix_mult (LTM_MATRIX (trans), LATTICE_BASE (lattice), LTM_MATRIX (trans1), depth, depth, depth); /* Compute the Hermite normal form for the new transformation matrix. */ H = lambda_trans_matrix_new (depth, depth); U = lambda_trans_matrix_new (depth, depth); lambda_matrix_hermite (LTM_MATRIX (trans1), depth, LTM_MATRIX (H), LTM_MATRIX (U)); /* Compute the auxiliary loop nest's space from the unimodular portion. */ auxillary_nest = lambda_compute_auxillary_space (nest, U, lambda_obstack); /* Compute the loop step signs from the old step signs and the transformation matrix. */ stepsigns = lambda_compute_step_signs (trans1, stepsigns); /* Compute the target loop nest space from the auxiliary nest and the lower triangular matrix H. */ target_nest = lambda_compute_target_space (auxillary_nest, H, stepsigns, lambda_obstack); origin = lambda_vector_new (depth); origin_invariants = lambda_matrix_new (depth, invariants); lambda_matrix_vector_mult (LTM_MATRIX (trans), depth, depth, LATTICE_ORIGIN (lattice), origin); lambda_matrix_mult (LTM_MATRIX (trans), LATTICE_ORIGIN_INVARIANTS (lattice), origin_invariants, depth, depth, invariants); for (i = 0; i < depth; i++) { loop = LN_LOOPS (target_nest)[i]; expression = LL_LINEAR_OFFSET (loop); if (lambda_vector_zerop (LLE_COEFFICIENTS (expression), depth)) f = 1; else f = LLE_DENOMINATOR (expression); LLE_CONSTANT (expression) += f * origin[i]; for (j = 0; j < invariants; j++) LLE_INVARIANT_COEFFICIENTS (expression)[j] += f * origin_invariants[i][j]; } return target_nest; } /* Convert a gcc tree expression EXPR to a lambda linear expression, and return the new expression. DEPTH is the depth of the loopnest. OUTERINDUCTIONVARS is an array of the induction variables for outer loops in this nest. INVARIANTS is the array of invariants for the loop. EXTRA is the amount we have to add/subtract from the expression because of the type of comparison it is used in. */ static lambda_linear_expression gcc_tree_to_linear_expression (int depth, tree expr, VEC(tree,heap) *outerinductionvars, VEC(tree,heap) *invariants, int extra, struct obstack * lambda_obstack) { lambda_linear_expression lle = NULL; switch (TREE_CODE (expr)) { case INTEGER_CST: { lle = lambda_linear_expression_new (depth, 2 * depth, lambda_obstack); LLE_CONSTANT (lle) = TREE_INT_CST_LOW (expr); if (extra != 0) LLE_CONSTANT (lle) += extra; LLE_DENOMINATOR (lle) = 1; } break; case SSA_NAME: { tree iv, invar; size_t i; for (i = 0; VEC_iterate (tree, outerinductionvars, i, iv); i++) if (iv != NULL) { if (SSA_NAME_VAR (iv) == SSA_NAME_VAR (expr)) { lle = lambda_linear_expression_new (depth, 2 * depth, lambda_obstack); LLE_COEFFICIENTS (lle)[i] = 1; if (extra != 0) LLE_CONSTANT (lle) = extra; LLE_DENOMINATOR (lle) = 1; } } for (i = 0; VEC_iterate (tree, invariants, i, invar); i++) if (invar != NULL) { if (SSA_NAME_VAR (invar) == SSA_NAME_VAR (expr)) { lle = lambda_linear_expression_new (depth, 2 * depth, lambda_obstack); LLE_INVARIANT_COEFFICIENTS (lle)[i] = 1; if (extra != 0) LLE_CONSTANT (lle) = extra; LLE_DENOMINATOR (lle) = 1; } } } break; default: return NULL; } return lle; } /* Return the depth of the loopnest NEST */ static int depth_of_nest (struct loop *nest) { size_t depth = 0; while (nest) { depth++; nest = nest->inner; } return depth; } /* Return true if OP is invariant in LOOP and all outer loops. */ static bool invariant_in_loop_and_outer_loops (struct loop *loop, tree op) { if (is_gimple_min_invariant (op)) return true; if (loop_depth (loop) == 0) return true; if (!expr_invariant_in_loop_p (loop, op)) return false; if (!invariant_in_loop_and_outer_loops (loop_outer (loop), op)) return false; return true; } /* Generate a lambda loop from a gcc loop LOOP. Return the new lambda loop, or NULL if it could not be converted. DEPTH is the depth of the loop. INVARIANTS is a pointer to the array of loop invariants. The induction variable for this loop should be stored in the parameter OURINDUCTIONVAR. OUTERINDUCTIONVARS is an array of induction variables for outer loops. */ static lambda_loop gcc_loop_to_lambda_loop (struct loop *loop, int depth, VEC(tree,heap) ** invariants, tree * ourinductionvar, VEC(tree,heap) * outerinductionvars, VEC(tree,heap) ** lboundvars, VEC(tree,heap) ** uboundvars, VEC(int,heap) ** steps, struct obstack * lambda_obstack) { gimple phi; gimple exit_cond; tree access_fn, inductionvar; tree step; lambda_loop lloop = NULL; lambda_linear_expression lbound, ubound; tree test_lhs, test_rhs; int stepint; int extra = 0; tree lboundvar, uboundvar, uboundresult; /* Find out induction var and exit condition. */ inductionvar = find_induction_var_from_exit_cond (loop); exit_cond = get_loop_exit_condition (loop); if (inductionvar == NULL || exit_cond == NULL) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot determine exit condition or induction variable for loop.\n"); return NULL; } if (SSA_NAME_DEF_STMT (inductionvar) == NULL) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot find PHI node for induction variable\n"); return NULL; } phi = SSA_NAME_DEF_STMT (inductionvar); if (gimple_code (phi) != GIMPLE_PHI) { tree op = SINGLE_SSA_TREE_OPERAND (phi, SSA_OP_USE); if (!op) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot find PHI node for induction variable\n"); return NULL; } phi = SSA_NAME_DEF_STMT (op); if (gimple_code (phi) != GIMPLE_PHI) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot find PHI node for induction variable\n"); return NULL; } } /* The induction variable name/version we want to put in the array is the result of the induction variable phi node. */ *ourinductionvar = PHI_RESULT (phi); access_fn = instantiate_parameters (loop, analyze_scalar_evolution (loop, PHI_RESULT (phi))); if (access_fn == chrec_dont_know) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Access function for induction variable phi is unknown\n"); return NULL; } step = evolution_part_in_loop_num (access_fn, loop->num); if (!step || step == chrec_dont_know) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot determine step of loop.\n"); return NULL; } if (TREE_CODE (step) != INTEGER_CST) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Step of loop is not integer.\n"); return NULL; } stepint = TREE_INT_CST_LOW (step); /* Only want phis for induction vars, which will have two arguments. */ if (gimple_phi_num_args (phi) != 2) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: PHI node for induction variable has >2 arguments\n"); return NULL; } /* Another induction variable check. One argument's source should be in the loop, one outside the loop. */ if (flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, 0)->src) && flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, 1)->src)) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: PHI edges both inside loop, or both outside loop.\n"); return NULL; } if (flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, 0)->src)) { lboundvar = PHI_ARG_DEF (phi, 1); lbound = gcc_tree_to_linear_expression (depth, lboundvar, outerinductionvars, *invariants, 0, lambda_obstack); } else { lboundvar = PHI_ARG_DEF (phi, 0); lbound = gcc_tree_to_linear_expression (depth, lboundvar, outerinductionvars, *invariants, 0, lambda_obstack); } if (!lbound) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot convert lower bound to linear expression\n"); return NULL; } /* One part of the test may be a loop invariant tree. */ VEC_reserve (tree, heap, *invariants, 1); test_lhs = gimple_cond_lhs (exit_cond); test_rhs = gimple_cond_rhs (exit_cond); if (TREE_CODE (test_rhs) == SSA_NAME && invariant_in_loop_and_outer_loops (loop, test_rhs)) VEC_quick_push (tree, *invariants, test_rhs); else if (TREE_CODE (test_lhs) == SSA_NAME && invariant_in_loop_and_outer_loops (loop, test_lhs)) VEC_quick_push (tree, *invariants, test_lhs); /* The non-induction variable part of the test is the upper bound variable. */ if (test_lhs == inductionvar) uboundvar = test_rhs; else uboundvar = test_lhs; /* We only size the vectors assuming we have, at max, 2 times as many invariants as we do loops (one for each bound). This is just an arbitrary number, but it has to be matched against the code below. */ gcc_assert (VEC_length (tree, *invariants) <= (unsigned int) (2 * depth)); /* We might have some leftover. */ if (gimple_cond_code (exit_cond) == LT_EXPR) extra = -1 * stepint; else if (gimple_cond_code (exit_cond) == NE_EXPR) extra = -1 * stepint; else if (gimple_cond_code (exit_cond) == GT_EXPR) extra = -1 * stepint; else if (gimple_cond_code (exit_cond) == EQ_EXPR) extra = 1 * stepint; ubound = gcc_tree_to_linear_expression (depth, uboundvar, outerinductionvars, *invariants, extra, lambda_obstack); uboundresult = build2 (PLUS_EXPR, TREE_TYPE (uboundvar), uboundvar, build_int_cst (TREE_TYPE (uboundvar), extra)); VEC_safe_push (tree, heap, *uboundvars, uboundresult); VEC_safe_push (tree, heap, *lboundvars, lboundvar); VEC_safe_push (int, heap, *steps, stepint); if (!ubound) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Unable to convert loop: Cannot convert upper bound to linear expression\n"); return NULL; } lloop = lambda_loop_new (); LL_STEP (lloop) = stepint; LL_LOWER_BOUND (lloop) = lbound; LL_UPPER_BOUND (lloop) = ubound; return lloop; } /* Given a LOOP, find the induction variable it is testing against in the exit condition. Return the induction variable if found, NULL otherwise. */ tree find_induction_var_from_exit_cond (struct loop *loop) { gimple expr = get_loop_exit_condition (loop); tree ivarop; tree test_lhs, test_rhs; if (expr == NULL) return NULL_TREE; if (gimple_code (expr) != GIMPLE_COND) return NULL_TREE; test_lhs = gimple_cond_lhs (expr); test_rhs = gimple_cond_rhs (expr); /* Find the side that is invariant in this loop. The ivar must be the other side. */ if (expr_invariant_in_loop_p (loop, test_lhs)) ivarop = test_rhs; else if (expr_invariant_in_loop_p (loop, test_rhs)) ivarop = test_lhs; else return NULL_TREE; if (TREE_CODE (ivarop) != SSA_NAME) return NULL_TREE; return ivarop; } DEF_VEC_P(lambda_loop); DEF_VEC_ALLOC_P(lambda_loop,heap); /* Generate a lambda loopnest from a gcc loopnest LOOP_NEST. Return the new loop nest. INDUCTIONVARS is a pointer to an array of induction variables for the loopnest that will be filled in during this process. INVARIANTS is a pointer to an array of invariants that will be filled in during this process. */ lambda_loopnest gcc_loopnest_to_lambda_loopnest (struct loop *loop_nest, VEC(tree,heap) **inductionvars, VEC(tree,heap) **invariants, struct obstack * lambda_obstack) { lambda_loopnest ret = NULL; struct loop *temp = loop_nest; int depth = depth_of_nest (loop_nest); size_t i; VEC(lambda_loop,heap) *loops = NULL; VEC(tree,heap) *uboundvars = NULL; VEC(tree,heap) *lboundvars = NULL; VEC(int,heap) *steps = NULL; lambda_loop newloop; tree inductionvar = NULL; bool perfect_nest = perfect_nest_p (loop_nest); if (!perfect_nest && !can_convert_to_perfect_nest (loop_nest)) goto fail; while (temp) { newloop = gcc_loop_to_lambda_loop (temp, depth, invariants, &inductionvar, *inductionvars, &lboundvars, &uboundvars, &steps, lambda_obstack); if (!newloop) goto fail; VEC_safe_push (tree, heap, *inductionvars, inductionvar); VEC_safe_push (lambda_loop, heap, loops, newloop); temp = temp->inner; } if (!perfect_nest) { if (!perfect_nestify (loop_nest, lboundvars, uboundvars, steps, *inductionvars)) { if (dump_file) fprintf (dump_file, "Not a perfect loop nest and couldn't convert to one.\n"); goto fail; } else if (dump_file) fprintf (dump_file, "Successfully converted loop nest to perfect loop nest.\n"); } ret = lambda_loopnest_new (depth, 2 * depth, lambda_obstack); for (i = 0; VEC_iterate (lambda_loop, loops, i, newloop); i++) LN_LOOPS (ret)[i] = newloop; fail: VEC_free (lambda_loop, heap, loops); VEC_free (tree, heap, uboundvars); VEC_free (tree, heap, lboundvars); VEC_free (int, heap, steps); return ret; } /* Convert a lambda body vector LBV to a gcc tree, and return the new tree. STMTS_TO_INSERT is a pointer to a tree where the statements we need to be inserted for us are stored. INDUCTION_VARS is the array of induction variables for the loop this LBV is from. TYPE is the tree type to use for the variables and trees involved. */ static tree lbv_to_gcc_expression (lambda_body_vector lbv, tree type, VEC(tree,heap) *induction_vars, gimple_seq *stmts_to_insert) { int k; tree resvar; tree expr = build_linear_expr (type, LBV_COEFFICIENTS (lbv), induction_vars); k = LBV_DENOMINATOR (lbv); gcc_assert (k != 0); if (k != 1) expr = fold_build2 (CEIL_DIV_EXPR, type, expr, build_int_cst (type, k)); resvar = create_tmp_var (type, "lbvtmp"); add_referenced_var (resvar); return force_gimple_operand (fold (expr), stmts_to_insert, true, resvar); } /* Convert a linear expression from coefficient and constant form to a gcc tree. Return the tree that represents the final value of the expression. LLE is the linear expression to convert. OFFSET is the linear offset to apply to the expression. TYPE is the tree type to use for the variables and math. INDUCTION_VARS is a vector of induction variables for the loops. INVARIANTS is a vector of the loop nest invariants. WRAP specifies what tree code to wrap the results in, if there is more than one (it is either MAX_EXPR, or MIN_EXPR). STMTS_TO_INSERT Is a pointer to the statement list we fill in with statements that need to be inserted for the linear expression. */ static tree lle_to_gcc_expression (lambda_linear_expression lle, lambda_linear_expression offset, tree type, VEC(tree,heap) *induction_vars, VEC(tree,heap) *invariants, enum tree_code wrap, gimple_seq *stmts_to_insert) { int k; tree resvar; tree expr = NULL_TREE; VEC(tree,heap) *results = NULL; gcc_assert (wrap == MAX_EXPR || wrap == MIN_EXPR); /* Build up the linear expressions. */ for (; lle != NULL; lle = LLE_NEXT (lle)) { expr = build_linear_expr (type, LLE_COEFFICIENTS (lle), induction_vars); expr = fold_build2 (PLUS_EXPR, type, expr, build_linear_expr (type, LLE_INVARIANT_COEFFICIENTS (lle), invariants)); k = LLE_CONSTANT (lle); if (k) expr = fold_build2 (PLUS_EXPR, type, expr, build_int_cst (type, k)); k = LLE_CONSTANT (offset); if (k) expr = fold_build2 (PLUS_EXPR, type, expr, build_int_cst (type, k)); k = LLE_DENOMINATOR (lle); if (k != 1) expr = fold_build2 (wrap == MAX_EXPR ? CEIL_DIV_EXPR : FLOOR_DIV_EXPR, type, expr, build_int_cst (type, k)); expr = fold (expr); VEC_safe_push (tree, heap, results, expr); } gcc_assert (expr); /* We may need to wrap the results in a MAX_EXPR or MIN_EXPR. */ if (VEC_length (tree, results) > 1) { size_t i; tree op; expr = VEC_index (tree, results, 0); for (i = 1; VEC_iterate (tree, results, i, op); i++) expr = fold_build2 (wrap, type, expr, op); } VEC_free (tree, heap, results); resvar = create_tmp_var (type, "lletmp"); add_referenced_var (resvar); return force_gimple_operand (fold (expr), stmts_to_insert, true, resvar); } /* Remove the induction variable defined at IV_STMT. */ void remove_iv (gimple iv_stmt) { gimple_stmt_iterator si = gsi_for_stmt (iv_stmt); if (gimple_code (iv_stmt) == GIMPLE_PHI) { unsigned i; for (i = 0; i < gimple_phi_num_args (iv_stmt); i++) { gimple stmt; imm_use_iterator imm_iter; tree arg = gimple_phi_arg_def (iv_stmt, i); bool used = false; if (TREE_CODE (arg) != SSA_NAME) continue; FOR_EACH_IMM_USE_STMT (stmt, imm_iter, arg) if (stmt != iv_stmt && !is_gimple_debug (stmt)) used = true; if (!used) remove_iv (SSA_NAME_DEF_STMT (arg)); } remove_phi_node (&si, true); } else { gsi_remove (&si, true); release_defs (iv_stmt); } } /* Transform a lambda loopnest NEW_LOOPNEST, which had TRANSFORM applied to it, back into gcc code. This changes the loops, their induction variables, and their bodies, so that they match the transformed loopnest. OLD_LOOPNEST is the loopnest before we've replaced it with the new loopnest. OLD_IVS is a vector of induction variables from the old loopnest. INVARIANTS is a vector of loop invariants from the old loopnest. NEW_LOOPNEST is the new lambda loopnest to replace OLD_LOOPNEST with. TRANSFORM is the matrix transform that was applied to OLD_LOOPNEST to get NEW_LOOPNEST. */ void lambda_loopnest_to_gcc_loopnest (struct loop *old_loopnest, VEC(tree,heap) *old_ivs, VEC(tree,heap) *invariants, VEC(gimple,heap) **remove_ivs, lambda_loopnest new_loopnest, lambda_trans_matrix transform, struct obstack * lambda_obstack) { struct loop *temp; size_t i = 0; unsigned j; size_t depth = 0; VEC(tree,heap) *new_ivs = NULL; tree oldiv; gimple_stmt_iterator bsi; transform = lambda_trans_matrix_inverse (transform); if (dump_file) { fprintf (dump_file, "Inverse of transformation matrix:\n"); print_lambda_trans_matrix (dump_file, transform); } depth = depth_of_nest (old_loopnest); temp = old_loopnest; while (temp) { lambda_loop newloop; basic_block bb; edge exit; tree ivvar, ivvarinced; gimple exitcond; gimple_seq stmts; enum tree_code testtype; tree newupperbound, newlowerbound; lambda_linear_expression offset; tree type; bool insert_after; gimple inc_stmt; oldiv = VEC_index (tree, old_ivs, i); type = TREE_TYPE (oldiv); /* First, build the new induction variable temporary */ ivvar = create_tmp_var (type, "lnivtmp"); add_referenced_var (ivvar); VEC_safe_push (tree, heap, new_ivs, ivvar); newloop = LN_LOOPS (new_loopnest)[i]; /* Linear offset is a bit tricky to handle. Punt on the unhandled cases for now. */ offset = LL_LINEAR_OFFSET (newloop); gcc_assert (LLE_DENOMINATOR (offset) == 1 && lambda_vector_zerop (LLE_COEFFICIENTS (offset), depth)); /* Now build the new lower bounds, and insert the statements necessary to generate it on the loop preheader. */ stmts = NULL; newlowerbound = lle_to_gcc_expression (LL_LOWER_BOUND (newloop), LL_LINEAR_OFFSET (newloop), type, new_ivs, invariants, MAX_EXPR, &stmts); if (stmts) { gsi_insert_seq_on_edge (loop_preheader_edge (temp), stmts); gsi_commit_edge_inserts (); } /* Build the new upper bound and insert its statements in the basic block of the exit condition */ stmts = NULL; newupperbound = lle_to_gcc_expression (LL_UPPER_BOUND (newloop), LL_LINEAR_OFFSET (newloop), type, new_ivs, invariants, MIN_EXPR, &stmts); exit = single_exit (temp); exitcond = get_loop_exit_condition (temp); bb = gimple_bb (exitcond); bsi = gsi_after_labels (bb); if (stmts) gsi_insert_seq_before (&bsi, stmts, GSI_NEW_STMT); /* Create the new iv. */ standard_iv_increment_position (temp, &bsi, &insert_after); create_iv (newlowerbound, build_int_cst (type, LL_STEP (newloop)), ivvar, temp, &bsi, insert_after, &ivvar, NULL); /* Unfortunately, the incremented ivvar that create_iv inserted may not dominate the block containing the exit condition. So we simply create our own incremented iv to use in the new exit test, and let redundancy elimination sort it out. */ inc_stmt = gimple_build_assign_with_ops (PLUS_EXPR, SSA_NAME_VAR (ivvar), ivvar, build_int_cst (type, LL_STEP (newloop))); ivvarinced = make_ssa_name (SSA_NAME_VAR (ivvar), inc_stmt); gimple_assign_set_lhs (inc_stmt, ivvarinced); bsi = gsi_for_stmt (exitcond); gsi_insert_before (&bsi, inc_stmt, GSI_SAME_STMT); /* Replace the exit condition with the new upper bound comparison. */ testtype = LL_STEP (newloop) >= 0 ? LE_EXPR : GE_EXPR; /* We want to build a conditional where true means exit the loop, and false means continue the loop. So swap the testtype if this isn't the way things are.*/ if (exit->flags & EDGE_FALSE_VALUE) testtype = swap_tree_comparison (testtype); gimple_cond_set_condition (exitcond, testtype, newupperbound, ivvarinced); update_stmt (exitcond); VEC_replace (tree, new_ivs, i, ivvar); i++; temp = temp->inner; } /* Rewrite uses of the old ivs so that they are now specified in terms of the new ivs. */ for (i = 0; VEC_iterate (tree, old_ivs, i, oldiv); i++) { imm_use_iterator imm_iter; use_operand_p use_p; tree oldiv_def; gimple oldiv_stmt = SSA_NAME_DEF_STMT (oldiv); gimple stmt; if (gimple_code (oldiv_stmt) == GIMPLE_PHI) oldiv_def = PHI_RESULT (oldiv_stmt); else oldiv_def = SINGLE_SSA_TREE_OPERAND (oldiv_stmt, SSA_OP_DEF); gcc_assert (oldiv_def != NULL_TREE); FOR_EACH_IMM_USE_STMT (stmt, imm_iter, oldiv_def) { tree newiv; gimple_seq stmts; lambda_body_vector lbv, newlbv; if (is_gimple_debug (stmt)) continue; /* Compute the new expression for the induction variable. */ depth = VEC_length (tree, new_ivs); lbv = lambda_body_vector_new (depth, lambda_obstack); LBV_COEFFICIENTS (lbv)[i] = 1; newlbv = lambda_body_vector_compute_new (transform, lbv, lambda_obstack); stmts = NULL; newiv = lbv_to_gcc_expression (newlbv, TREE_TYPE (oldiv), new_ivs, &stmts); if (stmts && gimple_code (stmt) != GIMPLE_PHI) { bsi = gsi_for_stmt (stmt); gsi_insert_seq_before (&bsi, stmts, GSI_SAME_STMT); } FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) propagate_value (use_p, newiv); if (stmts && gimple_code (stmt) == GIMPLE_PHI) for (j = 0; j < gimple_phi_num_args (stmt); j++) if (gimple_phi_arg_def (stmt, j) == newiv) gsi_insert_seq_on_edge (gimple_phi_arg_edge (stmt, j), stmts); update_stmt (stmt); } /* Remove the now unused induction variable. */ VEC_safe_push (gimple, heap, *remove_ivs, oldiv_stmt); } VEC_free (tree, heap, new_ivs); } /* Return TRUE if this is not interesting statement from the perspective of determining if we have a perfect loop nest. */ static bool not_interesting_stmt (gimple stmt) { /* Note that COND_EXPR's aren't interesting because if they were exiting the loop, we would have already failed the number of exits tests. */ if (gimple_code (stmt) == GIMPLE_LABEL || gimple_code (stmt) == GIMPLE_GOTO || gimple_code (stmt) == GIMPLE_COND || is_gimple_debug (stmt)) return true; return false; } /* Return TRUE if PHI uses DEF for it's in-the-loop edge for LOOP. */ static bool phi_loop_edge_uses_def (struct loop *loop, gimple phi, tree def) { unsigned i; for (i = 0; i < gimple_phi_num_args (phi); i++) if (flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src)) if (PHI_ARG_DEF (phi, i) == def) return true; return false; } /* Return TRUE if STMT is a use of PHI_RESULT. */ static bool stmt_uses_phi_result (gimple stmt, tree phi_result) { tree use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE); /* This is conservatively true, because we only want SIMPLE bumpers of the form x +- constant for our pass. */ return (use == phi_result); } /* STMT is a bumper stmt for LOOP if the version it defines is used in the in-loop-edge in a phi node, and the operand it uses is the result of that phi node. I.E. i_29 = i_3 + 1 i_3 = PHI (0, i_29); */ static bool stmt_is_bumper_for_loop (struct loop *loop, gimple stmt) { gimple use; tree def; imm_use_iterator iter; use_operand_p use_p; def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF); if (!def) return false; FOR_EACH_IMM_USE_FAST (use_p, iter, def) { use = USE_STMT (use_p); if (gimple_code (use) == GIMPLE_PHI) { if (phi_loop_edge_uses_def (loop, use, def)) if (stmt_uses_phi_result (stmt, PHI_RESULT (use))) return true; } } return false; } /* Return true if LOOP is a perfect loop nest. Perfect loop nests are those loop nests where all code occurs in the innermost loop body. If S is a program statement, then i.e. DO I = 1, 20 S1 DO J = 1, 20 ... END DO END DO is not a perfect loop nest because of S1. DO I = 1, 20 DO J = 1, 20 S1 ... END DO END DO is a perfect loop nest. Since we don't have high level loops anymore, we basically have to walk our statements and ignore those that are there because the loop needs them (IE the induction variable increment, and jump back to the top of the loop). */ bool perfect_nest_p (struct loop *loop) { basic_block *bbs; size_t i; gimple exit_cond; /* Loops at depth 0 are perfect nests. */ if (!loop->inner) return true; bbs = get_loop_body (loop); exit_cond = get_loop_exit_condition (loop); for (i = 0; i < loop->num_nodes; i++) { if (bbs[i]->loop_father == loop) { gimple_stmt_iterator bsi; for (bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi); gsi_next (&bsi)) { gimple stmt = gsi_stmt (bsi); if (gimple_code (stmt) == GIMPLE_COND && exit_cond != stmt) goto non_perfectly_nested; if (stmt == exit_cond || not_interesting_stmt (stmt) || stmt_is_bumper_for_loop (loop, stmt)) continue; non_perfectly_nested: free (bbs); return false; } } } free (bbs); return perfect_nest_p (loop->inner); } /* Replace the USES of X in STMT, or uses with the same step as X with Y. YINIT is the initial value of Y, REPLACEMENTS is a hash table to avoid creating duplicate temporaries and FIRSTBSI is statement iterator where new temporaries should be inserted at the beginning of body basic block. */ static void replace_uses_equiv_to_x_with_y (struct loop *loop, gimple stmt, tree x, int xstep, tree y, tree yinit, htab_t replacements, gimple_stmt_iterator *firstbsi) { ssa_op_iter iter; use_operand_p use_p; FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) { tree use = USE_FROM_PTR (use_p); tree step = NULL_TREE; tree scev, init, val, var; gimple setstmt; struct tree_map *h, in; void **loc; /* Replace uses of X with Y right away. */ if (use == x) { SET_USE (use_p, y); continue; } scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, use)); if (scev == NULL || scev == chrec_dont_know) continue; step = evolution_part_in_loop_num (scev, loop->num); if (step == NULL || step == chrec_dont_know || TREE_CODE (step) != INTEGER_CST || int_cst_value (step) != xstep) continue; /* Use REPLACEMENTS hash table to cache already created temporaries. */ in.hash = htab_hash_pointer (use); in.base.from = use; h = (struct tree_map *) htab_find_with_hash (replacements, &in, in.hash); if (h != NULL) { SET_USE (use_p, h->to); continue; } /* USE which has the same step as X should be replaced with a temporary set to Y + YINIT - INIT. */ init = initial_condition_in_loop_num (scev, loop->num); gcc_assert (init != NULL && init != chrec_dont_know); if (TREE_TYPE (use) == TREE_TYPE (y)) { val = fold_build2 (MINUS_EXPR, TREE_TYPE (y), init, yinit); val = fold_build2 (PLUS_EXPR, TREE_TYPE (y), y, val); if (val == y) { /* If X has the same type as USE, the same step and same initial value, it can be replaced by Y. */ SET_USE (use_p, y); continue; } } else { val = fold_build2 (MINUS_EXPR, TREE_TYPE (y), y, yinit); val = fold_convert (TREE_TYPE (use), val); val = fold_build2 (PLUS_EXPR, TREE_TYPE (use), val, init); } /* Create a temporary variable and insert it at the beginning of the loop body basic block, right after the PHI node which sets Y. */ var = create_tmp_var (TREE_TYPE (use), "perfecttmp"); add_referenced_var (var); val = force_gimple_operand_gsi (firstbsi, val, false, NULL, true, GSI_SAME_STMT); setstmt = gimple_build_assign (var, val); var = make_ssa_name (var, setstmt); gimple_assign_set_lhs (setstmt, var); gsi_insert_before (firstbsi, setstmt, GSI_SAME_STMT); update_stmt (setstmt); SET_USE (use_p, var); h = GGC_NEW (struct tree_map); h->hash = in.hash; h->base.from = use; h->to = var; loc = htab_find_slot_with_hash (replacements, h, in.hash, INSERT); gcc_assert ((*(struct tree_map **)loc) == NULL); *(struct tree_map **) loc = h; } } /* Return true if STMT is an exit PHI for LOOP */ static bool exit_phi_for_loop_p (struct loop *loop, gimple stmt) { if (gimple_code (stmt) != GIMPLE_PHI || gimple_phi_num_args (stmt) != 1 || gimple_bb (stmt) != single_exit (loop)->dest) return false; return true; } /* Return true if STMT can be put back into the loop INNER, by copying it to the beginning of that loop and changing the uses. */ static bool can_put_in_inner_loop (struct loop *inner, gimple stmt) { imm_use_iterator imm_iter; use_operand_p use_p; gcc_assert (is_gimple_assign (stmt)); if (gimple_vuse (stmt) || !stmt_invariant_in_loop_p (inner, stmt)) return false; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_assign_lhs (stmt)) { if (!exit_phi_for_loop_p (inner, USE_STMT (use_p))) { basic_block immbb = gimple_bb (USE_STMT (use_p)); if (!flow_bb_inside_loop_p (inner, immbb)) return false; } } return true; } /* Return true if STMT can be put *after* the inner loop of LOOP. */ static bool can_put_after_inner_loop (struct loop *loop, gimple stmt) { imm_use_iterator imm_iter; use_operand_p use_p; if (gimple_vuse (stmt)) return false; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_assign_lhs (stmt)) { if (!exit_phi_for_loop_p (loop, USE_STMT (use_p))) { basic_block immbb = gimple_bb (USE_STMT (use_p)); if (!dominated_by_p (CDI_DOMINATORS, immbb, loop->inner->header) && !can_put_in_inner_loop (loop->inner, stmt)) return false; } } return true; } /* Return true when the induction variable IV is simple enough to be re-synthesized. */ static bool can_duplicate_iv (tree iv, struct loop *loop) { tree scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, iv)); if (!automatically_generated_chrec_p (scev)) { tree step = evolution_part_in_loop_num (scev, loop->num); if (step && step != chrec_dont_know && TREE_CODE (step) == INTEGER_CST) return true; } return false; } /* If this is a scalar operation that can be put back into the inner loop, or after the inner loop, through copying, then do so. This works on the theory that any amount of scalar code we have to reduplicate into or after the loops is less expensive that the win we get from rearranging the memory walk the loop is doing so that it has better cache behavior. */ static bool cannot_convert_modify_to_perfect_nest (gimple stmt, struct loop *loop) { use_operand_p use_a, use_b; imm_use_iterator imm_iter; ssa_op_iter op_iter, op_iter1; tree op0 = gimple_assign_lhs (stmt); /* The statement should not define a variable used in the inner loop. */ if (TREE_CODE (op0) == SSA_NAME && !can_duplicate_iv (op0, loop)) FOR_EACH_IMM_USE_FAST (use_a, imm_iter, op0) if (gimple_bb (USE_STMT (use_a))->loop_father == loop->inner) return true; FOR_EACH_SSA_USE_OPERAND (use_a, stmt, op_iter, SSA_OP_USE) { gimple node; tree op = USE_FROM_PTR (use_a); /* The variables should not be used in both loops. */ if (!can_duplicate_iv (op, loop)) FOR_EACH_IMM_USE_FAST (use_b, imm_iter, op) if (gimple_bb (USE_STMT (use_b))->loop_father == loop->inner) return true; /* The statement should not use the value of a scalar that was modified in the loop. */ node = SSA_NAME_DEF_STMT (op); if (gimple_code (node) == GIMPLE_PHI) FOR_EACH_PHI_ARG (use_b, node, op_iter1, SSA_OP_USE) { tree arg = USE_FROM_PTR (use_b); if (TREE_CODE (arg) == SSA_NAME) { gimple arg_stmt = SSA_NAME_DEF_STMT (arg); if (gimple_bb (arg_stmt) && (gimple_bb (arg_stmt)->loop_father == loop->inner)) return true; } } } return false; } /* Return true when BB contains statements that can harm the transform to a perfect loop nest. */ static bool cannot_convert_bb_to_perfect_nest (basic_block bb, struct loop *loop) { gimple_stmt_iterator bsi; gimple exit_condition = get_loop_exit_condition (loop); for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) { gimple stmt = gsi_stmt (bsi); if (stmt == exit_condition || not_interesting_stmt (stmt) || stmt_is_bumper_for_loop (loop, stmt)) continue; if (is_gimple_assign (stmt)) { if (cannot_convert_modify_to_perfect_nest (stmt, loop)) return true; if (can_duplicate_iv (gimple_assign_lhs (stmt), loop)) continue; if (can_put_in_inner_loop (loop->inner, stmt) || can_put_after_inner_loop (loop, stmt)) continue; } /* If the bb of a statement we care about isn't dominated by the header of the inner loop, then we can't handle this case right now. This test ensures that the statement comes completely *after* the inner loop. */ if (!dominated_by_p (CDI_DOMINATORS, gimple_bb (stmt), loop->inner->header)) return true; } return false; } /* Return TRUE if LOOP is an imperfect nest that we can convert to a perfect one. At the moment, we only handle imperfect nests of depth 2, where all of the statements occur after the inner loop. */ static bool can_convert_to_perfect_nest (struct loop *loop) { basic_block *bbs; size_t i; gimple_stmt_iterator si; /* Can't handle triply nested+ loops yet. */ if (!loop->inner || loop->inner->inner) return false; bbs = get_loop_body (loop); for (i = 0; i < loop->num_nodes; i++) if (bbs[i]->loop_father == loop && cannot_convert_bb_to_perfect_nest (bbs[i], loop)) goto fail; /* We also need to make sure the loop exit only has simple copy phis in it, otherwise we don't know how to transform it into a perfect nest. */ for (si = gsi_start_phis (single_exit (loop)->dest); !gsi_end_p (si); gsi_next (&si)) if (gimple_phi_num_args (gsi_stmt (si)) != 1) goto fail; free (bbs); return true; fail: free (bbs); return false; } DEF_VEC_I(source_location); DEF_VEC_ALLOC_I(source_location,heap); /* Transform the loop nest into a perfect nest, if possible. LOOP is the loop nest to transform into a perfect nest LBOUNDS are the lower bounds for the loops to transform UBOUNDS are the upper bounds for the loops to transform STEPS is the STEPS for the loops to transform. LOOPIVS is the induction variables for the loops to transform. Basically, for the case of FOR (i = 0; i < 50; i++) { FOR (j =0; j < 50; j++) { <whatever> } <some code> } This function will transform it into a perfect loop nest by splitting the outer loop into two loops, like so: FOR (i = 0; i < 50; i++) { FOR (j = 0; j < 50; j++) { <whatever> } } FOR (i = 0; i < 50; i ++) { <some code> } Return FALSE if we can't make this loop into a perfect nest. */ static bool perfect_nestify (struct loop *loop, VEC(tree,heap) *lbounds, VEC(tree,heap) *ubounds, VEC(int,heap) *steps, VEC(tree,heap) *loopivs) { basic_block *bbs; gimple exit_condition; gimple cond_stmt; basic_block preheaderbb, headerbb, bodybb, latchbb, olddest; int i; gimple_stmt_iterator bsi, firstbsi; bool insert_after; edge e; struct loop *newloop; gimple phi; tree uboundvar; gimple stmt; tree oldivvar, ivvar, ivvarinced; VEC(tree,heap) *phis = NULL; VEC(source_location,heap) *locations = NULL; htab_t replacements = NULL; /* Create the new loop. */ olddest = single_exit (loop)->dest; preheaderbb = split_edge (single_exit (loop)); headerbb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb); /* Push the exit phi nodes that we are moving. */ for (bsi = gsi_start_phis (olddest); !gsi_end_p (bsi); gsi_next (&bsi)) { phi = gsi_stmt (bsi); VEC_reserve (tree, heap, phis, 2); VEC_reserve (source_location, heap, locations, 1); VEC_quick_push (tree, phis, PHI_RESULT (phi)); VEC_quick_push (tree, phis, PHI_ARG_DEF (phi, 0)); VEC_quick_push (source_location, locations, gimple_phi_arg_location (phi, 0)); } e = redirect_edge_and_branch (single_succ_edge (preheaderbb), headerbb); /* Remove the exit phis from the old basic block. */ for (bsi = gsi_start_phis (olddest); !gsi_end_p (bsi); ) remove_phi_node (&bsi, false); /* and add them back to the new basic block. */ while (VEC_length (tree, phis) != 0) { tree def; tree phiname; source_location locus; def = VEC_pop (tree, phis); phiname = VEC_pop (tree, phis); locus = VEC_pop (source_location, locations); phi = create_phi_node (phiname, preheaderbb); add_phi_arg (phi, def, single_pred_edge (preheaderbb), locus); } flush_pending_stmts (e); VEC_free (tree, heap, phis); bodybb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb); latchbb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb); make_edge (headerbb, bodybb, EDGE_FALLTHRU); cond_stmt = gimple_build_cond (NE_EXPR, integer_one_node, integer_zero_node, NULL_TREE, NULL_TREE); bsi = gsi_start_bb (bodybb); gsi_insert_after (&bsi, cond_stmt, GSI_NEW_STMT); e = make_edge (bodybb, olddest, EDGE_FALSE_VALUE); make_edge (bodybb, latchbb, EDGE_TRUE_VALUE); make_edge (latchbb, headerbb, EDGE_FALLTHRU); /* Update the loop structures. */ newloop = duplicate_loop (loop, olddest->loop_father); newloop->header = headerbb; newloop->latch = latchbb; add_bb_to_loop (latchbb, newloop); add_bb_to_loop (bodybb, newloop); add_bb_to_loop (headerbb, newloop); set_immediate_dominator (CDI_DOMINATORS, bodybb, headerbb); set_immediate_dominator (CDI_DOMINATORS, headerbb, preheaderbb); set_immediate_dominator (CDI_DOMINATORS, preheaderbb, single_exit (loop)->src); set_immediate_dominator (CDI_DOMINATORS, latchbb, bodybb); set_immediate_dominator (CDI_DOMINATORS, olddest, recompute_dominator (CDI_DOMINATORS, olddest)); /* Create the new iv. */ oldivvar = VEC_index (tree, loopivs, 0); ivvar = create_tmp_var (TREE_TYPE (oldivvar), "perfectiv"); add_referenced_var (ivvar); standard_iv_increment_position (newloop, &bsi, &insert_after); create_iv (VEC_index (tree, lbounds, 0), build_int_cst (TREE_TYPE (oldivvar), VEC_index (int, steps, 0)), ivvar, newloop, &bsi, insert_after, &ivvar, &ivvarinced); /* Create the new upper bound. This may be not just a variable, so we copy it to one just in case. */ exit_condition = get_loop_exit_condition (newloop); uboundvar = create_tmp_var (TREE_TYPE (VEC_index (tree, ubounds, 0)), "uboundvar"); add_referenced_var (uboundvar); stmt = gimple_build_assign (uboundvar, VEC_index (tree, ubounds, 0)); uboundvar = make_ssa_name (uboundvar, stmt); gimple_assign_set_lhs (stmt, uboundvar); if (insert_after) gsi_insert_after (&bsi, stmt, GSI_SAME_STMT); else gsi_insert_before (&bsi, stmt, GSI_SAME_STMT); update_stmt (stmt); gimple_cond_set_condition (exit_condition, GE_EXPR, uboundvar, ivvarinced); update_stmt (exit_condition); replacements = htab_create_ggc (20, tree_map_hash, tree_map_eq, NULL); bbs = get_loop_body_in_dom_order (loop); /* Now move the statements, and replace the induction variable in the moved statements with the correct loop induction variable. */ oldivvar = VEC_index (tree, loopivs, 0); firstbsi = gsi_start_bb (bodybb); for (i = loop->num_nodes - 1; i >= 0 ; i--) { gimple_stmt_iterator tobsi = gsi_last_bb (bodybb); if (bbs[i]->loop_father == loop) { /* If this is true, we are *before* the inner loop. If this isn't true, we are *after* it. The only time can_convert_to_perfect_nest returns true when we have statements before the inner loop is if they can be moved into the inner loop. The only time can_convert_to_perfect_nest returns true when we have statements after the inner loop is if they can be moved into the new split loop. */ if (dominated_by_p (CDI_DOMINATORS, loop->inner->header, bbs[i])) { gimple_stmt_iterator header_bsi = gsi_after_labels (loop->inner->header); for (bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi);) { gimple stmt = gsi_stmt (bsi); if (stmt == exit_condition || not_interesting_stmt (stmt) || stmt_is_bumper_for_loop (loop, stmt)) { gsi_next (&bsi); continue; } gsi_move_before (&bsi, &header_bsi); } } else { /* Note that the bsi only needs to be explicitly incremented when we don't move something, since it is automatically incremented when we do. */ for (bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi);) { gimple stmt = gsi_stmt (bsi); if (stmt == exit_condition || not_interesting_stmt (stmt) || stmt_is_bumper_for_loop (loop, stmt)) { gsi_next (&bsi); continue; } replace_uses_equiv_to_x_with_y (loop, stmt, oldivvar, VEC_index (int, steps, 0), ivvar, VEC_index (tree, lbounds, 0), replacements, &firstbsi); gsi_move_before (&bsi, &tobsi); /* If the statement has any virtual operands, they may need to be rewired because the original loop may still reference them. */ if (gimple_vuse (stmt)) mark_sym_for_renaming (gimple_vop (cfun)); } } } } free (bbs); htab_delete (replacements); return perfect_nest_p (loop); } /* Return true if TRANS is a legal transformation matrix that respects the dependence vectors in DISTS and DIRS. The conservative answer is false. "Wolfe proves that a unimodular transformation represented by the matrix T is legal when applied to a loop nest with a set of lexicographically non-negative distance vectors RDG if and only if for each vector d in RDG, (T.d >= 0) is lexicographically positive. i.e.: if and only if it transforms the lexicographically positive distance vectors to lexicographically positive vectors. Note that a unimodular matrix must transform the zero vector (and only it) to the zero vector." S.Muchnick. */ bool lambda_transform_legal_p (lambda_trans_matrix trans, int nb_loops, VEC (ddr_p, heap) *dependence_relations) { unsigned int i, j; lambda_vector distres; struct data_dependence_relation *ddr; gcc_assert (LTM_COLSIZE (trans) == nb_loops && LTM_ROWSIZE (trans) == nb_loops); /* When there are no dependences, the transformation is correct. */ if (VEC_length (ddr_p, dependence_relations) == 0) return true; ddr = VEC_index (ddr_p, dependence_relations, 0); if (ddr == NULL) return true; /* When there is an unknown relation in the dependence_relations, we know that it is no worth looking at this loop nest: give up. */ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) return false; distres = lambda_vector_new (nb_loops); /* For each distance vector in the dependence graph. */ for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++) { /* Don't care about relations for which we know that there is no dependence, nor about read-read (aka. output-dependences): these data accesses can happen in any order. */ if (DDR_ARE_DEPENDENT (ddr) == chrec_known || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr)))) continue; /* Conservatively answer: "this transformation is not valid". */ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) return false; /* If the dependence could not be captured by a distance vector, conservatively answer that the transform is not valid. */ if (DDR_NUM_DIST_VECTS (ddr) == 0) return false; /* Compute trans.dist_vect */ for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++) { lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops, DDR_DIST_VECT (ddr, j), distres); if (!lambda_vector_lexico_pos (distres, nb_loops)) return false; } } return true; } /* Collects parameters from affine function ACCESS_FUNCTION, and push them in PARAMETERS. */ static void lambda_collect_parameters_from_af (tree access_function, struct pointer_set_t *param_set, VEC (tree, heap) **parameters) { if (access_function == NULL) return; if (TREE_CODE (access_function) == SSA_NAME && pointer_set_contains (param_set, access_function) == 0) { pointer_set_insert (param_set, access_function); VEC_safe_push (tree, heap, *parameters, access_function); } else { int i, num_operands = tree_operand_length (access_function); for (i = 0; i < num_operands; i++) lambda_collect_parameters_from_af (TREE_OPERAND (access_function, i), param_set, parameters); } } /* Collects parameters from DATAREFS, and push them in PARAMETERS. */ void lambda_collect_parameters (VEC (data_reference_p, heap) *datarefs, VEC (tree, heap) **parameters) { unsigned i, j; struct pointer_set_t *parameter_set = pointer_set_create (); data_reference_p data_reference; for (i = 0; VEC_iterate (data_reference_p, datarefs, i, data_reference); i++) for (j = 0; j < DR_NUM_DIMENSIONS (data_reference); j++) lambda_collect_parameters_from_af (DR_ACCESS_FN (data_reference, j), parameter_set, parameters); pointer_set_destroy (parameter_set); } /* Translates BASE_EXPR to vector CY. AM is needed for inferring indexing positions in the data access vector. CST is the analyzed integer constant. */ static bool av_for_af_base (tree base_expr, lambda_vector cy, struct access_matrix *am, int cst) { bool result = true; switch (TREE_CODE (base_expr)) { case INTEGER_CST: /* Constant part. */ cy[AM_CONST_COLUMN_INDEX (am)] += int_cst_value (base_expr) * cst; return true; case SSA_NAME: { int param_index = access_matrix_get_index_for_parameter (base_expr, am); if (param_index >= 0) { cy[param_index] = cst + cy[param_index]; return true; } return false; } case PLUS_EXPR: return av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, cst) && av_for_af_base (TREE_OPERAND (base_expr, 1), cy, am, cst); case MINUS_EXPR: return av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, cst) && av_for_af_base (TREE_OPERAND (base_expr, 1), cy, am, -1 * cst); case MULT_EXPR: if (TREE_CODE (TREE_OPERAND (base_expr, 0)) == INTEGER_CST) result = av_for_af_base (TREE_OPERAND (base_expr, 1), cy, am, cst * int_cst_value (TREE_OPERAND (base_expr, 0))); else if (TREE_CODE (TREE_OPERAND (base_expr, 1)) == INTEGER_CST) result = av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, cst * int_cst_value (TREE_OPERAND (base_expr, 1))); else result = false; return result; case NEGATE_EXPR: return av_for_af_base (TREE_OPERAND (base_expr, 0), cy, am, -1 * cst); default: return false; } return result; } /* Translates ACCESS_FUN to vector CY. AM is needed for inferring indexing positions in the data access vector. */ static bool av_for_af (tree access_fun, lambda_vector cy, struct access_matrix *am) { switch (TREE_CODE (access_fun)) { case POLYNOMIAL_CHREC: { tree left = CHREC_LEFT (access_fun); tree right = CHREC_RIGHT (access_fun); unsigned var; if (TREE_CODE (right) != INTEGER_CST) return false; var = am_vector_index_for_loop (am, CHREC_VARIABLE (access_fun)); cy[var] = int_cst_value (right); if (TREE_CODE (left) == POLYNOMIAL_CHREC) return av_for_af (left, cy, am); else return av_for_af_base (left, cy, am, 1); } case INTEGER_CST: /* Constant part. */ return av_for_af_base (access_fun, cy, am, 1); default: return false; } } /* Initializes the access matrix for DATA_REFERENCE. */ static bool build_access_matrix (data_reference_p data_reference, VEC (tree, heap) *parameters, VEC (loop_p, heap) *nest) { struct access_matrix *am = GGC_NEW (struct access_matrix); unsigned i, ndim = DR_NUM_DIMENSIONS (data_reference); unsigned nivs = VEC_length (loop_p, nest); unsigned lambda_nb_columns; AM_LOOP_NEST (am) = nest; AM_NB_INDUCTION_VARS (am) = nivs; AM_PARAMETERS (am) = parameters; lambda_nb_columns = AM_NB_COLUMNS (am); AM_MATRIX (am) = VEC_alloc (lambda_vector, gc, ndim); for (i = 0; i < ndim; i++) { lambda_vector access_vector = lambda_vector_new (lambda_nb_columns); tree access_function = DR_ACCESS_FN (data_reference, i); if (!av_for_af (access_function, access_vector, am)) return false; VEC_quick_push (lambda_vector, AM_MATRIX (am), access_vector); } DR_ACCESS_MATRIX (data_reference) = am; return true; } /* Returns false when one of the access matrices cannot be built. */ bool lambda_compute_access_matrices (VEC (data_reference_p, heap) *datarefs, VEC (tree, heap) *parameters, VEC (loop_p, heap) *nest) { data_reference_p dataref; unsigned ix; for (ix = 0; VEC_iterate (data_reference_p, datarefs, ix, dataref); ix++) if (!build_access_matrix (dataref, parameters, nest)) return false; return true; }
Go to most recent revision | Compare with Previous | Blame | View Log